TL;DR
This paper introduces the Epistemic Argumentation Framework (EAF), a novel way to integrate an agent's beliefs with abstract argumentation frameworks. EAFs allow us to model how an agent's subjective beliefs influence their reasoning about arguments, going beyond traditional frameworks that only consider objective evidence. This opens doors for more nuanced modeling of decision-making and agreement in multi-agent systems.
What Problem Does This Paper Address?
Traditional argumentation frameworks, like Dung's abstract argumentation frameworks (AFs), are excellent at representing arguments and their attacks, forming a basis for rational reasoning. However, they primarily focus on the objective structure of arguments and attacks. In real-world scenarios, rational agents don't just process objective evidence; their decisions are heavily influenced by their personal beliefs and biases. For instance, in a political debate or a consumer choice, an agent's pre-existing beliefs can shape how they interpret arguments and ultimately make a decision. Existing approaches struggle to distinguish between objective evidence and subjective beliefs, often merging them or making it difficult to represent dynamic changes in an agent's stance due to evolving beliefs. This paper aims to bridge this gap by providing a formal mechanism to explicitly represent and reason about an agent's beliefs within an argumentation framework.
Methods
The core contribution is the Epistemic Argumentation Framework (EAF), defined as a triple (AF, φ), where AF is a standard argumentation framework and φ is an epistemic constraint. This constraint is a formula expressed in an epistemic logic that captures the agent's beliefs about the argumentation framework. The semantics of an EAF are defined through w-epistemic labelling sets, which are collections of w-labellings (where w can be complete, stable, grounded, or preferred) that collectively satisfy the epistemic constraint.
Key methodological aspects include:
- Epistemic Logic: The paper utilizes an epistemic logic with operators like K (belief) and M (possibility) to express propositional formulas over the justification states of arguments (in, out, und).
- w-Epistemic Labelling Sets: These sets are the central semantic concept, representing the agent's belief state by filtering the possible labellings of the underlying AF.
- Representing Preferences: EAFs can represent preferences over arguments and even over justification states using specific epistemic constraints.
- Multi-Agent Systems: The framework is extended to model multiple agents sharing an AF but holding different beliefs, leading to concepts of credulous and skeptical agreement.
- Computation: The paper outlines an algorithm for checking the existence of non-empty w-epistemic labelling sets and discusses its complexity. It also proposes a transformation of EAFs into epistemic logic programs for computation.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Separation of Belief and Evidence: EAFs elegantly separate objective argumentation structure (AF) from subjective beliefs (φ), allowing for clearer analysis and easier belief revision.
- Belief Revision as Constraint Revision: An agent's changing beliefs can be modeled by simply revising the epistemic constraint φ, without altering the underlying argumentation framework. This is a powerful mechanism for dynamic reasoning.
- Modeling Nuanced Beliefs: The use of epistemic logic (K and M operators) allows for expressing not just what an agent knows to be true, but also what they believe to be possibly true, capturing a wider range of epistemic states.
- Preference as Epistemic Constraint: Preferences, a well-studied area in argumentation, are elegantly encoded as epistemic constraints, providing a unified perspective.
- Multi-Agent Agreement: EAFs provide a formal basis for understanding how agents with different beliefs can reach (or fail to reach) agreement on specific arguments.
Key Results
- The paper formally defines EAFs and their semantics using w-epistemic labelling sets.
- It demonstrates how EAFs can represent an agent's beliefs, preferences, and conditional preferences.
- The framework is extended to model multi-agent scenarios, defining notions of credulous and skeptical agreement.
- Complexity results are provided: checking for the existence of a non-empty grounded epistemic labelling set is P-complete, while for stable, complete, and preferred semantics, it is NP-complete.
- A method for transforming EAFs into epistemic logic programs for computation is presented.
Limitations and Future Work
- Complexity of Epistemic Formulas: While the paper provides complexity results, the expressiveness of epistemic formulas can lead to complex computations, especially when dealing with exponential numbers of disjuncts in DNF.
- Focus on Single Agent Beliefs: The primary focus is on an agent's own beliefs. Extending this to reasoning about beliefs of other agents and incorporating beliefs of others into one's own reasoning is identified as future work.
- Connection to Extension Enforcement: The problem of modifying an AF to ensure a desired set of arguments becomes a subset of an extension is related to EAFs and is left for future investigation.
- Probabilistic Argumentation Connection: The paper suggests a potential connection between EAFs and probabilistic argumentation, leaving the precise formulation and proof for future work.
- Structured Argumentation: The current framework is for abstract argumentation. Extending it to structured argumentation (e.g., defeasible argumentation) is a natural next step.
Summary updated